You are on page 1of 20

CACTI 6.

0: A Tool to Understand Large Caches


Naveen Muralimanohar, Rajeev Balasubramonian , Norman P. Jouppi School of Computing, University of Utah Hewlett-Packard Laboratories

Abstract
Future processors will likely have large on-chip caches with a possibility of dedicating an entire die for on-chip storage in a 3D stacked model. With the ever growing disparity between transistor and wire delay, the properties of such large caches will primarily depend on the characteristics of the interconnection networks that connect various sub-modules of a cache. CACTI 6.0 is a signicantly enhanced version of the tool that primarily focuses on interconnect design for large caches. In addition to strengthening the existing analytical model of the tool for dominant cache components, CACTI 6.0 includes two major extensions over earlier versions: First, ability to model Non-Uniform Cache Access (NUCA). Second, ability to model different types of wires, such as RC based wires with different power, delay, and area characteristics and differential low-swing buses. The report details the analytical model assumed for the newly added modules along with their validation analysis.

Contents
1 Background 2 CACTI Terminologies 3 New features in CACTI 6.0 4 NUCA Modeling 4.1 Interconnect Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5 Analytical Models 5.1 Wire Parasitics . . . . . . . 5.2 Global Wires . . . . . . . . 5.3 Low-swing Wires . . . . . . 5.3.1 Transmitter . . . . . 5.3.2 Differential Wires . . 5.3.3 Sense Amplier . . . 5.4. Router Models . . . . . . . . 5.5 Distributed Wordline Model 5.6 Distributed Bitline Model . . 6 Trade-off Analysis 7. Validation 8 Usage 9 Conclusions References 3 3 4 5 7 9 10 10 11 11 13 14 15 15 15 16 17 18 19 20

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

Bitlines

Input address

Data output bits

Tag array

Decoder

Data array

Wordline

Bank

Address bits

Column muxes Sense Amps

Comparators Mux drivers Output driver Data output Valid output? Output driver

(a) Logical organization of a cache.

(b) Example physical organization of the data array.

Figure 1. Logical and physical organization of the cache (from CACTI 3.0 [13]).

1 Background
This section presents some basics on the CACTI cache access model. Figure 1(a) shows the basic logical structure of a uniform cache access (UCA) organization. The address request to the cache is rst provided as input to the decoder, which then activates a wordline in the data array and tag array. The contents of an entire row (referred to as a set) are placed on the bitlines, which are then sensed. The multiple tags thus read out of the tag array are compared against the input address to detect if one of the ways of the set does contain the requested data. This comparator logic drives the multiplexor that nally forwards at most one of the ways read out of the data array back to the requesting processor. The CACTI cache access model [14] takes in the following major parameters as input: cache capacity, cache block size (also known as cache line size), cache associativity, technology generation, number of ports, and number of independent banks (not sharing address and data lines). As output, it produces the cache conguration that minimizes delay (with a few exceptions), along with its power and area characteristics. CACTI models the delay/power/area of eight major cache components: decoder, wordline, bitline, senseamp, comparator, multiplexor, output driver, and inter-bank wires. The wordline and bitline delays are two of the most signicant components of the access time. The wordline and bitline delays are quadratic functions of the width and height of each array, respectively. In practice, the tag and data arrays are large enough that it is inefcient to implement them as single large structures. Hence, CACTI partitions each storage array (in the horizontal and vertical dimensions) to produce smaller sub-arrays and reduce wordline and bitline delays. The bitline is partitioned into Ndbl different segments, the wordline is partitioned into Ndwl segments, and so on. Each sub-array has its own decoder and some central pre-decoding is now required to route the request to the correct sub-array. CACTI carries out an exhaustive search across different sub-array counts (different values of Ndbl, Ndwl, etc.) and sub-array aspect ratios to compute the cache organization with optimal total delay. Typically, the cache is organized into a handful of banks. An example of the caches physical structure is shown in Figure 1(b).

2 CACTI Terminologies
The following is a list of keywords introduced by various versions of CACTI. Bank - A memory structure that consists of a data and a tag array. A cache is typically split into multiple banks and CACTI assumes enough bandwidth so that these banks can be accessed simultaneously. The network topology that interconnects these banks can vary depending on the cache model (UCA or NUCA).

Sub-arrays - A data or tag array is divided into a number of sub-arrays to reduce the delay due to wordline and bitline. Unlike banks, at any given time, these sub-arrays support only one single access. The total number of sub-arrays in a cache is equal to the product of Ndwl and Ndbl. Mat - A group of four sub-arrays (2x2) that share a common central predecoder. CACTIs exhaustive search starts from a minimum of at least one mat. Sub-bank - In a typical cache, a cache block is scattered across multiple sub-arrays to improve the reliability of a cache. Irrespective of the cache organization, CACTI assumes that every cache block in a cache is distributed across an entire row of mats and the row number corresponding to a particular block is determined based on the block address. Each row (of mats) in an array is referred to as a sub-bank. Ntwl/Ndwl - Number of horizontal partitions in a tag or data array i.e., the number of segments that a single wordline is partitioned into. Ntbl/Ndbl - Number of vertical partitions in a tag or data array i.e., the number of segments that a single bitline is partitioned into. Ntspd/Nspd - Number of sets stored in each row of a sub-array. For a given Ndwl and Ndbl values, Nspd decides the aspect ratio of the sub-array. Ntcm/Ndcm - Degree of bitline multiplexing. Ntsam/Ndsam - Degree of sense-amplier multiplexing.

3 New features in CACTI 6.0


CACTI 6.0 comes with a number of new features, most of which are targeted to improve the tools ability to model large caches. Incorporation of many different wire models for the inter-bank network: local/intermediate/global wires, repeater sizing/spacing for optimal delay or power, low-swing differential wires. Incorporation of models for router components (buffers, crossbar, arbiter). Introduction of grid topologies for NUCA and a shared bus architecture for UCA with low-swing wires. An algorithm for design space exploration that models different grid layouts and estimates average bank and network latency. The design space exploration also considers different wire and router types. The introduction of empirical network contention models to estimate the impact of network conguration, bank cycle time, and workload on average cache access delay. An improved and more accurate wordline and bitline delay model. A validation analysis of all new circuit models: low-swing differential wires, distributed RC model for wordlines and bitlines within cache banks (router components have been validated elsewhere). An improved interface that enables trade-off analysis for latency, power, cycle time, and area.

4 NUCA Modeling
Earlier versions of CACTI assumed a Uniform Cache Access (UCA) model in which, the access time of a cache is determined by the delay to access the farthest sub-array. To enable pipelining, an H-tree network is employed to connect all the sub-arrays of a cache. For large caches, this uniform model can suffer from a very high hit latency. A more scalable approach for future large caches is to replace the H-tree bus with a packet-switched on-chip grid network. The latency for a bank is determined by the delay to route the request and response between the bank that contains the data and the cache controller. Such a NUCA model was rst proposed by Kim et al. [7] and has been the subject of many architectural evaluations. CACTI 6.0 builds upon this model and adopts the following algorithm to identify the optimal NUCA organization. The tool rst iterates over a number of bank organizations: the cache is partitioned into 2N banks (where N varies from 1 to 12); for each N , the banks are organized in a grid with 2M rows (where M varies from 0 to N ). For each bank organization, CACTI 5.0 is employed to determine the optimal sub-array partitioning for the cache within each bank. Each bank is associated with a router. The average delay for a cache access is computed by estimating the number of network hops to each bank, the wire delay encountered on each hop, and the cache access delay within each bank. We further assume that each traversal through a router takes up R cycles, where R is a user-specied input. Router pipelines can be designed in many ways: a four-stage pipeline is commonly advocated [4], and recently, speculative pipelines that take up three, two, and one pipeline stage have also been proposed [4, 8, 11]. While we give the user the option to pick an aggressive or conservative router, the tool defaults to employing a moderately aggressive router pipeline with three stages. The user also has the exibility to specify the operating frequency of the network (which defaults to 5 GHz). However, based on the process technology and the router model, the tool will calculate the maximum possible network frequency [11]. If the assumed frequency is greater than the maximum possible value, the tool will downgrade the network frequency to the maximum value. In the above NUCA model, more partitions lead to smaller delays (and power) within each bank, but greater delays (and power) on the network (because of the constant overheads associated with each router and decoder). Hence, the above design space exploration is required to estimate the cache partition that yields optimal delay or power. The above algorithm was recently proposed by Muralimanohar and Balasubramonian [9]. While the algorithm is guaranteed to nd the cache structure with the lowest possible delay or power, the bandwidth of the cache might still not be sufcient enough for a multi core processor model. To address this problem, CACTI 6.0 further extends this algorithm by modeling contention in the network in much greater detail. This contention model itself has two major components. If the cache is partitioned into many banks, there are more routers/links on the network and the probability of two packets conicting at a router decrease. Thus, a many-banked cache is more capable of meeting the bandwidth demands of a many-core system. Further, certain aspects of the cache access within a bank cannot be easily pipelined. The longest such delay within the cache access (typically the bitline and sense-amp delays) represents the cycle time of the bank it is the minimum delay between successive accesses to that bank. A many-banked cache has relatively small banks and a relatively low cycle time, allowing it to support a higher throughput and lower wait-times once a request is delivered to the bank. Both of these two components (lower contention at routers and lower contention at banks) tend to favor a many-banked system. This aspect is also included in estimating the average access time for a given cache conguration. To improve the search space of the NUCA model, CACTI 6.0 also explores different router types and wire types for the links between adjacent routers. The wires are modeled as low-swing differential wires as well as global wires with different repeater congurations to yield many points in the power/delay/area spectrum. The sizes of buffers and virtual channels within a router have a major inuence on router power consumption as well as router contention under heavy load. By varying the number of virtual channels

300 250

16-core 8-core 4-core

400 350 300 Total No. of Cycles Network Latency Bank access latency Network contention Cycles

Contention Cycles

200

Latency (cycles)
2 4 8 16 32 64

250 200 150 100 50

150 100 50 0 Bank Count

0 2 4 8 No. of Banks 16 32 64

(a) Total network contention value/access for CMPs with different NUCA organizations

(b) Optimal NUCA organization

Figure 2. NUCA design space exploration.


Fetch queue size Bimodal predictor size Level 2 predictor Branch mispredict penalty Dispatch and commit width Register le size L1 I-cache L2 cache L2 Block size I and D TLB Network topology No. of virtual channels 64 16K 16K entries at least 12 cycles 8 100 (int and fp, each) 32KB 2-way 32MB 8-way SNUCA 64B 128 entries, 8KB page size Grid 4 /physical channel Branch predictor Level 1 predictor BTB size Fetch width Issue queue size Re-order Buffer size L1 D-cache comb. of bimodal and 2-level 16K entries, history 12 16K sets, 2-way 8 (across up to 2 basic blocks) 60 (int and fp, each) 80 32KB 2-way set-associative, 3 cycles, 4-way word-interleaved 300 cycles for the rst chunk Virtual channel Credit based ow control

Memory latency Flow control mechanism Back pressure handling

Table 1. Simplescalar simulator parameters.

per physical channel and the number of buffers per virtual channel, we are able to achieve different points on the router power-delay trade-off curve. The contention values for each considered NUCA cache organization are empirically estimated for typical workloads and incorporated into CACTI 6.0 as look-up tables. For each of the grid topologies considered (for different values of N and M ), we simulated L2 requests originating from single-core, two-core, four-core, eight-core, and sixteen-core processors. Each core executes a mix of programs from the SPEC benchmark suite. We divide the benchmark set into four categories, as described in Table 2. For every CMP organization, we run four sets of simulations, corresponding to each benchmark set tabulated. The generated cache trafc is then modeled on a detailed network simulator with support for virtual channel ow control. Details of the architectural and network simulator are listed in Table 1. The contention value (averaged across the various workloads) at routers and banks is estimated for each network topology and bank cycle time. Based on the user-specied inputs, the appropriate contention values in the look-up table are taken into account during the design space exploration.
Memory intensive benchmarks L2/L3 latency sensitive benchmarks Half latency sensitive & half non-latency sensitive benchmarks Random benchmark set applu, fma3d, swim, lucas equake, gap, vpr, art ammp, apsi, art, bzip2, crafty, eon, equake, gcc ammp, applu, lucas, bzip2 crafy, mgrid, mesa, gcc Entire SPEC suite

Table 2. Benchmark sets

For a network with completely pipelined links and routers, these contention values are only a function of the router topology and bank cycle time and will not be affected by process technology or L2 cache size1 . If CACTI is being employed to compute an optimal L3 cache organization, the contention values will likely be much less because the L2 cache lters out many requests. To handle this case, we also computed the average contention values assuming a large 2 MB L1 cache and this is incorporated into the model as well. In summary, the network contention values are impacted by the following parameters: M , N , bank cycle time, number of cores, router conguration (VCs, buffers), size of preceding cache. We plan to continue augmenting the tool with empirical contention values for other relevant sets of workloads such as commercial, multi-threaded, and transactional benchmarks with signicant trafc from cache coherence. Figure 2(b) shows an example design space exploration for a 32 MB NUCA L2 cache while attempting to minimize latency. The X-axis shows the number of banks that the cache is partitioned into. For each point on the X-axis, many different bank organizations are considered and the organization with optimal delay (averaged across all banks) is nally represented on the graph. The Y-axis represents this optimal delay and it is further broken down to represent the contributing components: bank access time, link and router delay, router and bank contention. We observe that the optimal delay is experienced when the cache is organized as a 2 4 grid of 8 banks.

4.1 Interconnect Model


With shrinking process technologies, interconnect plays an increasingly important role in deciding the power and performance of large structures. In the deep sub-micron era, the properties of a large cache are heavily impacted by the choice of the interconnect model [9, 10]. Another major enhancement to the tool that signicantly improves the search space is the inclusion of different wire models with varying power and delay characteristics. The properties of wires depend on a number of factors like dimensions, signaling, operating voltage, operating frequency, etc. Based on the signaling strategy, RC wires can be classied into two broad categories 2 : 1. Traditional full-swing wires, 2. Differential, low-swing, low power wires. The delay of an RC wire increases quadratically with its length. To avoid this quadratic relationship, a long wire is typically interleaved with repeaters at regular intervals. This makes delay a liner function of wire length. However, the use of repeaters at regular intervals requires that voltage levels of these wires swing across the full range (0-Vdd) for proper operation. Given the quadratic dependence between voltage and power, these full swing wires are accompanied by very high power overhead. Figure 5 shows the delay and power values of global wires for different process technologies. With power emerging as a major bottleneck, focusing singularly on performance is not possible. Alternatively, we can improve the power characteristics of these wires by incurring a delay penalty. In a typical, long, full swing wire, repeaters are one of the major contributors of interconnect power. Figure 4a shows the impact of repeater sizing and spacing on wire delay. Figure 4b, shows the contours corresponding to the 2% delay penalty increments for different repeater congurations. As we can see, by tolerating a delay penalty, signicant reduction in repeater overhead is possible. Figure 5 shows the power values of different wires that take 10%, 20%, and 30% delay penalty for different process technologies.
We assume here that the cache is organized as static-NUCA (SNUCA), where the address index bits determine the unique bank where the address can be found and the access distribution does not vary greatly as a function of the cache size. CACTI is designed to be more generic than specic. The contention values are provided as a guideline to most users. If a user is interested in a more specic NUCA policy, there is no substitute to generating the corresponding contention values and incorporating them in the tool. 2 Many recent proposals advocate designing wires with very low resistance and/or high operating frequency so that wires behave like a transmission line. While transmission lines incur very low delay, they are accompanied by high area overheads and suffer from signal integrity issues. For these reasons, we limit our discussion in this report to RC wires.
1

(b) Design space exploration with full-swing global wires (red, bottom (a) Design space exploration with global wires region), wires with 30% delay penalty (yellow, middle region), and differential low-swing wires (blue, top region)

Figure 3. Power/Delay trade-off in a 16MB UCA cache

(a) Effect repeater spacing/sizing on wire delay.

(b) Contours for 2% delay penalty. wires with 30% delay penalty (green), and differential low-swing wires (blue)

Figure 4. Repeater overhead vs Wire delay

(a) Delay characteristics of different wires at 32nm process (b) Energy characteristics of different wires at 32nm process technology. technology.

Figure 5. Power-delay properties of different wires

Figure 6. Interconnect segment

One of the primary reasons for the high power dissipation of global wires is the full swing requirement imposed by the repeaters. While we are able to somewhat reduce the power requirement by reducing repeater size and increasing repeater spacing, the requirement is still relatively high. Low voltage swing alternatives represent another mechanism to vary the wire power/delay/area trade-off. Reducing the voltage swing on global wires can result in a linear reduction in power. In addition, assuming a separate voltage source for low-swing drivers will result in a quadratic savings in power. But, these lucrative power savings are accompanied by many caveats. Since we can no longer use repeaters or latches, the delay of a low-swing wire increases quadratically with length. Since such a wire cannot be pipelined, they also suffer from lower throughput. A low-swing wire requires special transmitter and receiver circuits for signal generation and amplication. This not only increases the area requirement per bit, but also assigns a xed cost in terms of both delay and power for each bit traversal. In spite of these issues, the power savings possible through low-swing signalling makes it an attractive design choice. The detailed methodology for the design of low-swing wires and their overhead is described in a later section. In general, low-swing wires have superior power characteristics but incur high area and delay overheads. Figure 5 compares power delay characteristics of low-swing wires with global wires.

5 Analytical Models
The following sections discusses the analytical delay and power models for different wires. All the process specic parameters required for calculating the transistor and wire parasitics are obtained from ITRS [2].

9 9 9 9 0 0

0 0 0 0

E E E E E E

0 0 0 0 0 0

5 0 5 0 0 0

. . . . . .

2 2 1 1 5 0 e a y s D l ) (

5.1 Wire Parasitics


The resistance and capacitance per unit length of a wire is given by the following equations [5]: Rwire = d (thickness barrier)(width 2 barrier) (1)

where, d (< 1) is the loss in cross-sectional area due to dishing effect [2] and is the resistivity of the metal.

Cwire = thickness width + 2vert ) spacing layerspacing +f ringe(horiz , vert ) 0 (2Khoriz In the above equation for the capacitance, the rst term corresponds to the side wall capacitance, the second term models the capacitance due to wires in adjacent layers, and the last term corresponds to the fringing capacitance between the sidewall and the substrate.

5.2 Global Wires


For a long repeated wire, the single pole time constant model for the interconnect fragment shown in gure 6 is given by, rs 1 (2) = ( rs (c0 + cp ) + c + rsc0 + 0.5rcl) l s In the above equation, c0 is the capacitance of the minimum sized repeater, cp is its output parasitic capacitance, rs is its output resistance, l is the length of the interconnect segment between repeaters and s is the size of the repeater normalized to the minimum value. The values of c0 , cp , and rs are constant for a given process technology. Wire parasitics Rwire and Cwire represent resistance and capacitance per unit length. The optimal repeater sizing and spacing values can be calculated by differentiating equation 2 with respect to s and l and equating it to zero. Loptimal = 2rs (c0 + cp ) Rwire Cwire rs Cwire Rwire c0 (3)

Soptimal =

(4)

The delay value calculated using the above Loptimal and Soptimal is guaranteed to have minimum value. The total power dissipated is the sum of three main components (equation 5) [3]. Ptotal = Pswitching + Pshortcircuit + Pleakage The dynamic and leakage components of the interconnect are computed using equations 7 and 9. Soptimal (cp + c0 ) + c) Loptimal Loptimal (5)

2 Pdynamic = VDD fclock (

(6) (7)

+(VDD Wmin ISC fclock loge 3)Soptimal

fclock is the operating frequency, Wmin is the minimum width of the transistor, ISC is the short-circuit current, and the value ( /Loptimal ) can be calculated from equation 2. 10

Figure 7. Low-swing transmitter (actual transmitter has two such circuits to feed the differential wires)

( )optimal = 2 rs c0 rc 1 + L

0.5 1 +

cp c0

3 Pleakage = VDD Ileak Wn Soptimal (9) 2 Ileak is the leakage current and Wn is the minimum width of the NMOS transistor. With the above equations, we can compute the delay and power for global and semi-global wires. Wires faster than global wires can be obtained by increasing the wire width and spacing between the wires. Wires whose repeater spacing and sizing are different from equation 3 and 4 will incur a delay penalty. For a given delay penalty, the power optimal repeater size and spacing can be obtained from the contour shown in the gure 4b. The actual calculation involves solving a set of differential equations [3].

5.3 Low-swing Wires


A low-swing interconnect system consists of three main components: (1) Transmitter that generates and drives the low-swing signal, (2) Twisted differential wires, and (3) Receiver amplier. 5.3.1 Transmitter

For the transmitter circuit, we employ the model proposed by Ho et al. [6] shown in gure 7. For an RC tree with a time constant , the delay of the circuit for an input with nite rise time is given by equation 10, delayr = tf . [log vth 2 vth ] + 2trise b(1 )/tf Vdd V dd (10)

where, tf is the time constant of the tree, vth is the threshold voltage of the transistor, trise is the rise time of the input signal, and b is the fraction of the input swing in which the output changes (we assume b to be 0.5). For falling input, the equation changes to delayf = tf . [log(1 11 vth 2 2tf all .b.vth )] + Vdd tf .Vdd (11)

n n e

i i l b a n e

(8)

where, tf all is the fall time of the input. For the falling input, we use a value of 0.4 for b [18]. To get a reasonable estimate of the initial input signal rise/fall time, we consider two inverters connected in series. Let d be the delay of the second inverter. The tf all and trise values for the initial input can be approximated to d tf all = 1 vth d trise = vth The total delay of the transmitter is given by, tdelay = nanddelay + inverterdelay + driverdelay (12)

Each gate in the above equation (nand, inverter, and driver) can be reduced to a simple RC tree. Later Horowitz approximation is applied to calculate the delay of each gate. The power consumed in different gates can be derived from the input and output parasitics of the transistors. NAND gate: The equivalent resistance and capacitance values of a NAND gate is given by, Req = 2 Rnmos Ceq = 2 CP drain + 1.5 CN drain + CL where CL is the load capacitance of the NAND gate and is equal to the input capacitance of the next gate. The value of CL is equal to IN Vsize (CP gate + CN gate ) where, IN Vsize is the size of the inverter whose calculation is discussed in the later part of the section. NOTE: The drain capacitance of a transistor is highly non-linear. In the above equation for Ceq , the effective drain capacitance of two nmos transistors connected in series is approximated to 1.5 times the drain capacitance of a single nmos transistor. nand = Req Ceq Using the nand and trise values in equation 11, nanddelay can be calculated. Power consumed by the NAND gate is given by, 2 Pnand = Ceq Vdd The fall time (tf all ) of the input signal to the next stage (NOT gate) is given by tf all = nanddelay ( 1 ) 1 vth

Driver: To increase the energy savings in low-swing model, we assume a separate low voltage source for driving low-swing differential wires. The size of these drivers depends on its load capacitance which in turn depends on the length of the wire. To calculate the size of the driver, we rst calculate the drive resistance of the nmos transistors for a xed desired rise time of eight F04. Rdrive = Wdr = Risetime CL ln(0.5)

Rm Wmin Rdrive In the above equation, CL is the sum of capacitance of the wire and input capacitance of the sense amplier. Rm is the drive resistance of a minimum sized nmos transistor and Wmin is the width of the minimum sized transistor. 12

From the Rdrive value, the actual width of the pmos transistor can be calculated 3 . NOTE: The driver resistance Rdrive , calculated above is valid only if the supply voltage is set to full Vdd . Since low-swing drivers employ a separate low voltage source, the actual drive resistance of these transistors will be greater than the pmos transistor of the same size driven by the full Vdd . Hence, the Rdrive value is multiplied with an adjustment factor RES ADJ to account for the poor driving capability of the pmos transistor. Based on the SPICE simulation, RES ADJ value is calculated to be 8.6. NOT gate: The size of the NOT gate is calculated by applying the method of logical effort. Consider the NAND gate connected to the NOT gate that drives a load of CL , where, CL is equal to the input capacitance of the driver. path ef f ort = CL CN gate + CP gate

The delay will be minimum when the effort in each stage is same. stage ef f ort = CN OTi n = IN Vsize = (4/3) path ef f ort (4/3) CL stage ef f ort

CN OTi n CCN gate + CP gate

Using the above inverter size, the equivalent resistance and the capacitance of the gate can be calculated. Req = Rpmos Ceq = CP drain + CN drain + CL where CL for the inverter is equal to (2CN gate ). not = Req Ceq Using the above not and tf all values, notdelay can be calculated. Power consumed by this NOT gate is given by, 2 Pnot = Ceq Vdd The rise time for the next stage is given by trise = 5.3.2 Differential Wires notdelay vth

To alleviate the high delay overhead of the un-repeated low-swing wires, similar to differential bitlines, we employ pre-emphasis and pre-equalization optimizations. In pre-emphasis, the drive voltage of the driver is maintained at higher voltage than low-swing voltage. By overdriving these wires, it takes only a fraction of time constant to develop the differential voltage. In pre-equalization, after a bit traversal, the differential wires are shorted to recycle the charge. Developing a differential voltage on a pre-equalized wires takes less time compared to the wires with opposite polarity.
3

In our model, we limit the transistor width to 100 times the minimum size.

13

Figure 8. Sense-amplier circuit

The following equations present the time constant and capacitance values of the segment that consist of low-swing drivers and wires.

tdriver = (Rdriver (Cwire + Cdrain + Rwire Cwire /2 + (Rdriver + Rwire ) Csenseamp )

The Cwire and Rwire in the above equation represents resistance and capacitance parasitics of the low-swing wire. Rdriver and Cdrain are resistance and drain capacitance of the driver transistors. The pre-equalization and pre-emphasis optimization brings down this time constant to 35% of the above value. The total capacitance of the low-swing segment is given by Cload = Cwire + 2 Cdrain + Csensea mp The dynamic energy due to charging and discharging of differential wires is given by, Cload .VoverDrive .Vlowswing For our evaluations we assume an overdrive voltage of 400mV and a low swing voltage of 100mV. 5.3.3 Sense Amplier

Figure 8 shows the cross coupled inverter sense amplier circuit used at the receiver. The delay and power values of the sense amplier are directly calculated from the SPICE simulation. Table 3 shows the simulated values for different process technologies. To calculate these values, the sense amplier load is set to twice the input capacitance of the minimum sized inverter.

14

o t i b

(13)

Technology 90nm 68nm 45nm 32nm

Delay (ps) 279 200 38 30

Energy (fJ) 14.7 5.7 2.7 2.16

Table 3. Sense-amplier delay and energy values for different process technologies.

Figure 9. RC model of a wordline

5.4. Router Models


There have been a number of router proposals in the literature with different levels of speculation and pipeline stages [4, 8, 11]. The number of pipeline stages for router is left as a user-specied input, defaulting to 3 cycles. Buffers, crossbars, and arbiters are the major contributors to the router power. CACTI 6.0s analytical power models for crossbars and arbiters is similar to the model employed in Orion toolkit [17]. Buffer power is modeled using CACTIs inbuilt RAM model.

5.5 Distributed Wordline Model


Figure 9 shows the wordline circuit and its equivalent RC model. Earlier versions of CACTI modeled the wordline wire as a single lumped RC tree. In process technologies where wire parasitics dominate, a distributed RC model of the type shown in the gure will signicantly improve the accuracy of the model. Let cw and rw be the resistance and capacitance values of the wire of length l where, l is the width of the memory cell. The time constant governing the above RC tree is given by = Rdr Cdr + n Rdr (cw + Cpg ) + rw (cw + Cpg ) n (n + 1) 2

where, Rdr - Resistance of the pmos transistor in the driver. Cdr - Sum of the drain capacitance of the pmos and nmos transistors in the driver. Cpg - Input gate capacitance of the pass transistor. n - Length of the wordline in terms of number of memory cells.

5.6 Distributed Bitline Model


Figure 10 shows the RC model of the bitline read path. The time constant of the RC tree is given by, 15

Figure 10. RC model of a bitline

= Rpd Cpd + (Rpass + Rpd ) Cpass +

(Rpd + Rpass + r n + Rbmux ) Cbmux +

(Rpd + Rpass ) c n + n (n + 1) r c/2

where, Rpd - Resistance of the pull down transistor in the latch Cpd - Drain capacitance of the pull down transistor in the latch Rpass - Resistance of the pass transistor Cpass - Drain capacitance of the pass transistor Rbmux - Resistance of the transistor in the bitline multiplexer Cbmux - Drain capacitance of the transistor in the bitline multiplexer n - Length of the bitline in terms of number of memory cells c - Capacitance of the bitline segment between two memory cells that include wire capacitance and the drain capacitance of the pass transistor r - Resistance of the wire connecting two pass transistors We follow a methodology similar to the one proposed in the original version of CACTI [18] to take into account the effect of nite rise time of wordline signal on the bitline delay.

6 Trade-off Analysis
The new version of the tool adopts the following default cost function to evaluate a cache organization (taking into account delay, leakage power, dynamic power, cycle time, and area): cost = acc time + min acc time dyn power Wdyn power + min dyn power leak power + Wleak power min leak power cycle time Wcycle time + min cycle time area Warea min area Wacc
time

The weights for each term (Wacc time , Wdyn power , Wleak power , Wcycle time , Warea ) indicate the relative importance of each term and these are specied by the user as input parameters in the conguration le: 16

5000 4500 4000 3500

140

CACTI 6.0

120

CACTI 6.0 SPICE

Delay in (ps s)

3000 2500 2000 1500 1000 500 0 1 2 3 4 5 6 7 8 9 10

Energy/access (fJ)

SPICE

100 80 60 40 20 0 1 2 3 4 5 6 7 8 9 10

Wire Length (mm)

Length (mm)

(a) Delay verication

(b) Energy verication

Figure 11. Low-swing model verication

-weight 100 20 20 10 10 The above default weights used by the tool reect the priority of these metrics in a typical modern design. In addition, the following default line in the input parameters species the users willingness to deviate from the optimal set of metrics: -deviate 1000 1000 1000 1000 1000

The above line dictates that we are willing to consider a cache organization where each metric, say the access time, deviates from the lowest possible access time by 1000%. Hence, this default set of input parameters species a largely unconstrained search space. The following input lines restrict the tool to identify a cache organization that yields least power while giving up at most 10% performance: -weight 0 -deviate 10 100 100 0 0 1000 1000 1000 1000

7. Validation
In this work, we mainly focus on validating the new modules added to the framework. This includes low-swing wires, router components, and improved bitline and wordline models. Since SPICE results depend on the model les for transistors, we rst discuss the technology modeling changes made to the recent version of CACTI (version 5) and later detail our methodology for validating the newly added components to CACTI 6.0. Earlier versions of CACTI (version one through four) assumed linear technology scaling for calculating cache parameters. All the power, delay, and area values are rst calculated for 800nm technology and the results are linearly scaled to the user specied process value. While this approach is reasonably accurate for old process technologies, it can introduce non-trivial error for deep sub-micron technologies (less than 90nm). This problem is xed in CACTI 5 [15] by adopting ITRS parameters for all calculations. The current version of CACTI supports four different process technologies (90nm, 65nm, 45nm, and 32nm) with process specic values obtained from ITRS. Though ITRS projections are invaluable for quick analytical estimates, SPICE validation requires technology model les with greater detail and ITRS values cannot be directly plugged in for SPICE verication. The only non-commercial data available publicly for this purpose for recent process technologies is the Predictive Technology Model (PTM) [1]. For our validation, we employ the HSPICE tool along with the PTM 65 nm model le for validating the newly added components. The simulated values obtained from HSPICE are compared against CACTI 6.0 analytical models that take PTM parameters as input 4 . The analytical delay and power calculations
The PTM parameters employed for verication can be directly used for CACTI simulations. Since most architectural and circuit studies rely on ITRS parameters, CACTI by default assumes ITRS values to maintain consistency.
4

17

1000

1000

100

100

Delay (ps)

Delay (ps)

10

CACTI 6.0 SPICE

10

CACTI 6.0 SPICE

1 Memory Cells
4 2 0 1 2 1 5 6 5 2 8 2 1

1 32 64 128 No. of Cells 256

(a) Wordline

(b) Bitline

Figure 12. Distributed wordline and bitline model verication

performed by the tool primarily depend on the resistance and capacitance parasitics of transistors. For our validation, the capacitance values of source, drain, and gate of n and p transistors are derived from the PTM technology model le. The threshold voltage and the on-resistance of the transistors are calculated using SPICE simulations. In addition to modeling the gate delay and wire delay of different components, our analytical model also considers the delay penalty incurred due to the nite rise time and fall time of an input signal [18]. Figure 11 (a) & (b) show the comparison of delay and power values of the differential, low-swing analytical models against SPICE values. As mentioned earlier, a low-swing wire model can be broken into three components: transmitter (that generates the low-swing signal), differential wires5 , and sense ampliers. The modeling details of each of these components are discussed in section 5.3. Though the analytical model employed in CACTI 6.0 dynamically calculates the driver size appropriate for a given wire length, for the wire length of our interest, it ends up using the maximum driver size (which is set to 100 times the minimum transistor size) to incur minimum delay overhead. Earlier versions of CACTI also had the problem of over estimating the delay and power values of the sense-amplier. CACTI 6.0 eliminates this problem by directly using the SPICE generated values for sense-amp power and delay. On an average, the low-swing wire models are veried to be within 12% of the SPICE values. The lumped RC model used in prior versions of CACTI for bitlines and wordlines are replaced with a more accurate distributed RC model in CACTI 6.0. Based on a detailed spice modeling of the bitline segment along with the memory cells, we found the difference between the old and new model to be around 11% at 130 nm technology. This difference can go up to 50% with shrinking process technologies as wire parasitics become the dominant factor compared to transistor capacitance [12]. Figure 12 (a) & (b) compare the distributed wordline and bitline delay values and the SPICE values. The length of the wordlines or bitlines (specied in terms of memory array size) are carefully picked to represent a wide range of cache sizes. On an average, the new analytical models for the distributed wordlines and bitlines are veried to be within 13% and 12% of SPICE generated values. Buffers, crossbars, and arbiters are the primary components in a router. CACTI 6.0 uses its scratch RAM model to calculate read/write power for router buffers. We employ Orions arbiter and crossbar model for calculating router power and these models have been validated by Wang et al. [16].

8 Usage
Prior versions of CACTI take cache parameters such as cache size, block size, associativity, and technology as command line arguments. In addition to supporting the command line input, CACTI 6.0
5

Delay and power values of low-swing driver is also reported as part of differential wires.

18

also employs a conguration le (cache.cfg) to enable user to describe the cache parameters in much greater detail. The following are the valid command line arguments in CACTI 6.0: C B A Tech NoBanks and / or -weight <delay> <dynamic> <leakage> <cycle> <area> and / or -deviate <delay> <dynamic> <leakage> <cycle> <area> C - Cache size in bytes B - Block size in bytes A - Associativity Tech - Process technology in microns or nano-meter NoBanks - No. of UCA banks Command line arguments are optional in CACTI 6.0 and a more comprehensive description is possible using the conguration le. Other non-standard parameters that can be specied in the cache.cfg le include, No. of read ports, write ports, read-write ports in a cache H-tree bus width Operating temperature (which is used for calculating the cache leakage value), Custom tag size (that can be used to model special structures like branch target buffer, cache directory, etc.) Cache access mode (fast - low access time but power hungry; sequential - high access time but low power; Normal - less aggressive in terms of both power and delay) Cache type (DRAM, SRAM or a simple scratch RAM such as register les that does not need the tag array) NUCA bank count (By default CACTI calculates the optimal bank count value. However, user can force the tool to use a particular NUCA bank count value) Number of cores Cache level - L2 or L3 (Core count and cache level are used to calculate the contention values for a NUCA model) Design objective (weight and deviate parameters for NUCA and UCA) More details on each of these parameters is provided in the default cache.cfg le that is provided with the distribution.

9 Conclusions
The report details major revisions to the CACTI cache modeling tool along with a detailed description of the analytical model for newly added components. Interconnect plays a major role in deciding the delay and power values of large caches and we extended CACTIs design space exploration to carefully consider many different implementation choices for the interconnect components, including different wire types, routers, signaling strategy, and contention modeling. We also add modeling support for a

19

wide range of NUCA caches. CACTI 6.0 identies a number of relevant design choices on the powerdelay-area curves. The estimates of CACTI 6.0 can differ from the estimates of CACTI 5.0 signicantly, especially when more fully exploring the power-delay trade-off space. CACTI 6.0 is able to identify cache congurations that can reduce power by a factor of three, while incurring a 25% delay penalty. We valide components of the tool against Spice simulations and show good agreement between analytical and transistor-level models.

References
[1] Arizona State University. Predictive technology model. [2] S. I. Association. International Technology Roadmap for Semiconductors 2005. http://public.itrs.net/Links/2005ITRS/Home2005.htm. [3] K. Banerjee and A. Mehrotra. A Power-optimal Repeater Insertion Methodology for Global Interconnects in Nanometer Designs. IEEE Transactions on Electron Devices, 49(11):20012007, November 2002. [4] W. Dally and B. Towles. Principles and Practices of Interconnection Networks. Morgan Kaufmann, 1st edition, 2003. [5] R. Ho, K. Mai, and M. Horowitz. The Future of Wires. Proceedings of the IEEE, Vol.89, No.4, April 2001. [6] R. Ho, K. Mai, and M. Horowitz. Managing Wire Scaling: A Circuit Prespective. Interconnect Technology Conference, pages 177179, June 2003. [7] C. Kim, D. Burger, and S. Keckler. An Adaptive, Non-Uniform Cache Structure for Wire-Dominated OnChip Caches. In Proceedings of ASPLOS-X, October 2002. [8] R. Mullins, A. West, and S. Moore. Low-Latency Virtual-Channel Routers for On-Chip Networks. In Proceedings of ISCA-31, May 2004. [9] N. Muralimanohar and R. Balasubramonian. Interconnect Design Considerations for Large NUCA Caches. In Proceedings of the 34th International Symposium on Computer Architecture (ISCA-34), June 2007. [10] N. Muralimanohar, R. Balasubramonian, and N. P. Jouppi. Optimizing NUCA Organizations and Wiring Alternatives for Large Caches With CACTI 6.0. In Proceedings of MICRO-40, 2007. [11] L.-S. Peh and W. Dally. A Delay Model and Speculative Architecture for Pipelined Routers. In Proceedings of HPCA-7, 2001. [12] J. M. Rabaey, A. Chandrakasan, and B. Nikolic. Digital Integrated Circuits - A Design Perspective. PrenticeHall, 2nd edition, 2002. [13] P. Shivakumar and N. P. Jouppi. CACTI 3.0: An Integrated Cache Timing, Power, and Area Model. Technical Report TN-2001/2, Compaq Western Research Laboratory, August 2001. [14] D. Tarjan, S. Thoziyoor, and N. Jouppi. CACTI 4.0. Technical Report HPL-2006-86, HP Laboratories, 2006. [15] S. Thoziyoor, N. Muralimanohar, and N. Jouppi. CACTI 5.0. Technical Report HPL-2007-167, HP Laboratories, 2007. [16] H.-S. Wang, L.-S. Peh, and S. Malik. Power-Driven Design of Router Microarchitectures in On-Chip Networks. In Proceedings of MICRO-36, December 2003. [17] H.-S. Wang, X. Zhu, L.-S. Peh, and S. Malik. Orion: A Power-Performance Simulator for Interconnection Networks. In Proceedings of MICRO-35, November 2002. [18] S. Wilton and N. Jouppi. An Enhanced Access and Cycle Time Model for On-Chip Caches. Technical Report TN-93/5, Compaq Western Research Lab, 1993.

20